On 17.01.24 15:44, Mark Kanda wrote:
On 1/9/24 8:25 AM, David Hildenbrand wrote:On 09.01.24 15:15, Daniel P. Berrangé wrote:On Tue, Jan 09, 2024 at 03:02:00PM +0100, David Hildenbrand wrote:On 08.01.24 19:40, Mark Kanda wrote:On 1/8/24 9:40 AM, David Hildenbrand wrote:On 08.01.24 16:10, Mark Kanda wrote:Refactor the memory prealloc threads support: - Make memset context a global qlist - Move the memset thread join/cleanup code to a separate routineThis is functionally equivalent and facilitates multiple memset contexts (used in a subsequent patch). Signed-off-by: Mark Kanda <mark.ka...@oracle.com> --- util/oslib-posix.c | 104 +++++++++++++++++++++++++++++---------------- 1 file changed, 68 insertions(+), 36 deletions(-) diff --git a/util/oslib-posix.c b/util/oslib-posix.c index e86fd64e09..293297ac6c 100644 --- a/util/oslib-posix.c +++ b/util/oslib-posix.c @@ -63,11 +63,15 @@ struct MemsetThread; +static QLIST_HEAD(, MemsetContext) memset_contexts = + QLIST_HEAD_INITIALIZER(memset_contexts); + typedef struct MemsetContext { bool all_threads_created; bool any_thread_failed; struct MemsetThread *threads; int num_threads; + QLIST_ENTRY(MemsetContext) next; } MemsetContext; struct MemsetThread { @@ -81,7 +85,7 @@ struct MemsetThread { typedef struct MemsetThread MemsetThread; /* used by sigbus_handler() */ -static MemsetContext *sigbus_memset_context; +static bool sigbus_memset_context; struct sigaction sigbus_oldact; static QemuMutex sigbus_mutex; @@ -295,13 +299,16 @@ static void sigbus_handler(int signal) #endif /* CONFIG_LINUX */ { int i; + MemsetContext *context; if (sigbus_memset_context) { - for (i = 0; i < sigbus_memset_context->num_threads; i++) { - MemsetThread *thread = &sigbus_memset_context->threads[i]; + QLIST_FOREACH(context, &memset_contexts, next) { + for (i = 0; i < context->num_threads; i++) { + MemsetThread *thread = &context->threads[i]; - if (qemu_thread_is_self(&thread->pgthread)) { - siglongjmp(thread->env, 1); + if (qemu_thread_is_self(&thread->pgthread)) { + siglongjmp(thread->env, 1); + } } } } @@ -417,14 +424,15 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages, bool use_madv_populate_write) { static gsize initialized = 0; - MemsetContext context = { - .num_threads = get_memset_num_threads(hpagesize, numpages, max_threads), - }; + MemsetContext *context = g_malloc0(sizeof(MemsetContext)); size_t numpages_per_thread, leftover; void *(*touch_fn)(void *); - int ret = 0, i = 0; + int i = 0; char *addr = area; + context->num_threads = + get_memset_num_threads(hpagesize, numpages, max_threads); + if (g_once_init_enter(&initialized)) { qemu_mutex_init(&page_mutex); qemu_cond_init(&page_cond); @@ -433,7 +441,7 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages, if (use_madv_populate_write) { /* Avoid creating a single thread for MADV_POPULATE_WRITE */ - if (context.num_threads == 1) { + if (context->num_threads == 1) { if (qemu_madvise(area, hpagesize * numpages, QEMU_MADV_POPULATE_WRITE)) { return -errno; @@ -445,49 +453,74 @@ static int touch_all_pages(char *area, size_t hpagesize, size_t numpages, touch_fn = do_touch_pages; } - context.threads = g_new0(MemsetThread, context.num_threads); - numpages_per_thread = numpages / context.num_threads; - leftover = numpages % context.num_threads; - for (i = 0; i < context.num_threads; i++) { - context.threads[i].addr = addr; - context.threads[i].numpages = numpages_per_thread + (i < leftover); - context.threads[i].hpagesize = hpagesize; - context.threads[i].context = &context; + context->threads = g_new0(MemsetThread, context->num_threads); + numpages_per_thread = numpages / context->num_threads; + leftover = numpages % context->num_threads; + for (i = 0; i < context->num_threads; i++) { + context->threads[i].addr = addr; + context->threads[i].numpages = numpages_per_thread + (i < leftover); + context->threads[i].hpagesize = hpagesize; + context->threads[i].context = context; if (tc) { - thread_context_create_thread(tc, &context.threads[i].pgthread, + thread_context_create_thread(tc, &context->threads[i].pgthread, "touch_pages", - touch_fn, &context.threads[i], + touch_fn, &context->threads[i], QEMU_THREAD_JOINABLE); } else { - qemu_thread_create(&context.threads[i].pgthread, "touch_pages", - touch_fn, &context.threads[i], + qemu_thread_create(&context->threads[i].pgthread, "touch_pages", + touch_fn, &context->threads[i], QEMU_THREAD_JOINABLE); } - addr += context.threads[i].numpages * hpagesize; + addr += context->threads[i].numpages * hpagesize; } + QLIST_INSERT_HEAD(&memset_contexts, context, next); + if (!use_madv_populate_write) { - sigbus_memset_context = &context; + sigbus_memset_context = true;Thanks David,Could we just use the sigbus handling alone and support parallel init only when using MADV_POPULATE_WRITE where we don't have to mess with signal handlers?Ideally, we're hoping to support this with earlier kernels which don't support MADV_POPULATE _WRITE. But, I will check to see if we really need it.That's around since Linux 5.14, so please try simplifying.Further, how do you changes interact with other callers of qemu_prealloc_mem(), like virtio-mem?I'm not familiar with the intricacies of virtio-mem, but the basic idea of this series is to *only* allow parallel init during the start up phase (while prealloc_init == false). Once we have parsed all the command line args, we set prealloc_init = true (wait_mem_prealloc_init()) which causes all subsequent calls to qemu_prealloc_mem() to perform initialization synchronously. So, I *think* this covers the virtio-mem use case. Am I missing something?Good, so this should likely not affect virtio-mem (which also ends up preallocating memory when loading from a vmstate). To make this all a bit clearer, what about the following to make this: * Optimize for MADV_POPULATE_WRITE. If we really need support for !MADV_POPULATE_WRITE, this is better added on top later. * Pass in via a parameter that the caller requests async handling. "bool async" should be good enough. Then, pass that only from the memory backend call, while QEMU is still initializing (we can find a way to make that work). * Provide a function that waits for any remaining async os-mem-prealloc activity. That is essentially "wait_mem_prealloc_init", but without the special internal flag handling. Further, I do wonder if we want to make that behavior configurable. For example, one might want to initialize backends sequentially using 16 threads max, instead of consuming multiple times 16 threads concurrently.Seems to me that no matter what parallelisation we use (within mem regions, or across mem regions, or a mix of both), we should never use more threads than there are host CPUs.Yes. It gets tricky with multipe NUMA nodes, though. And that's what's being requested here.Can we have a pool of threads sized per available host CPUs that QEMU can access. Then for each memory backend fire off a set of init jobs that get distributed to "appropriate" threads in the pool. By appropriate I mean threads with the same NUMA affinity as the memory backend. This would give parallelisation both within a single large memory region, as well as across memory regions. eg 4 host CPUs, 3 memory regions (1GB for 1st numa node, 1GB for 2nd numa node, and 1 GB with no numa affinity). If we spread the init work then we end up withI'll note that having a mixture of numa + no numa is not a common config we should try to optimize.1st thread gets 500MB to init from 1st memory region, and 250MB to init from 3rd memory region 2st thread gets 500MB to init from 1st memory region, and 250MB to init from 3rd memory region 3st thread gets 500MB to init from 2nd memory region, and 250MB to init from 3rd memory region 4th thread gets 500MB to init from 2nd memory region, and 250MB to init from 3rd memory regionWe do have prealloc contexts that are used to craft new CPUs with the right NUMA affinity. Prealloc contexts are set for memory backends. So extending prealloc context to limit/reuse threads and do the pooling might be possible.From a NUMA perspective, I feel prealloc contexts already supports what we want - i.e. we can set the number of prealloc threads and their node affinity per memory region. As such, unless you feel strongly about it, I'm inclined to leave additional enhancements (thread pool) as a future project.
I tend to agree. It's uncommon to have many memory backends target the same NUMA node such that you would want to use pooling.
Usually, you'd expect in most vNUMA setups to have, say, 1 memory backend per vNUMA node that is mapped to some NUMA nodes in the parent.
Using prealloc contexts you can make sure that preallocating each individual one gets the right CPU affinity. Setting the prealloc threads you can limit the number of threads used per memory backend.
> Since multiple NUMA node configs are the primary motivation for parallel init, I do wonder if it should *only* be used in conjunction with prealloc context. Perhaps we should have parallel init only enabled for memory regions with prealloc context?
That sounds reasonable. -- Cheers, David / dhildenb
